Technical Essay
The trust estimator was designed to combine learning with physical structure without making exaggerated claims about formal guarantees. The main edge-trust model used graph context so that the reliability of one message could be interpreted relative to neighboring claims and the current platoon topology. An EdgeMLP variant scored each edge independently and later served as an architectural ablation.
The learning objective combined binary cross-entropy on trust labels with a physics-guided regularizer. The physics term penalized assigning high trust to a packet that implied a large kinematic inconsistency. In simplified notation:
L = L_BCE + lambda_phys L_phys.
The physics residual was not called a proof of correctness. It was a training signal that biased the estimator away from messages inconsistent with short-horizon longitudinal dynamics.
Reading Note
Graph neural network literature, particularly work on graph convolution and relational inductive biases, supported the decision to represent communication as edges rather than flattening all vehicle states into a fixed vector. Vehicle-platoon communication is inherently relational: the same claimed acceleration may be plausible for one sender-receiver pair and implausible for another.
The project also drew a line between "physics-guided" and "physics-certified." The former uses equations and residuals to structure learning; the latter would require a formal guarantee that was not established here. This distinction was preserved in the manuscript to prevent a common overclaim in physics-informed machine learning papers.